Interpretation of kidney graft biopsies using the Banff classification is still heterogeneous. In this study, extreme gradient boosting classifiers learned from two large training datasets (n = 631 and 304 cases) where the "reference diagnoses" were not strictly defined following the Banff rules but from central reading by expert pathologists and further interpreted consensually by experienced transplant nephrologists, in light of the clinical context. In three external validation datasets (n = 3744, 589, and 360), the classifiers yielded a mean ROC curve AUC (95%CI) of: 0.97 (0.92-1.00), 0.97 (0.96-0.97), and 0.95 (0.93-0.97) for antibody-mediated rejection (ABMR); 0.94 (0.91-0.96), 0.94 (0.92-0.95), and 0.91 (0.88-0.95) for T cell-mediated rejection; >0.96 (0.90-1.00) with all three for interstitial fibrosis-tubular atrophy. We also developed a classifier to discriminate active and chronic active ABMR with 95% accuracy. In conclusion, we built highly sensitive and specific artificial intelligence classifiers able to interpret kidney graft scoring together with a few clinical data and automatically diagnose rejection, with excellent concordance with the Banff rules and reference diagnoses made by a group of experts. Some discrepancies may point toward possible improvements that could be made to the Banff classification.
Keywords: biopsy; classification systems: Banff classification; clinical research / practice; kidney transplantation / nephrology; rejection: T cell mediated (TCMR); rejection: antibody-mediated (ABMR).
© 2022 The American Society of Transplantation and the American Society of Transplant Surgeons.